Hook
Last Thursday, my desk received a two-stage analysis report that had consumed eleven minutes of compute across nine valuation dimensions. The output contained one usable token: N/A, repeated roughly four hundred times. No technical read. No supply schedule. No regulatory assessment. Just a pipeline that had been handed an empty payload and โ against every instinct trained into a modern language model โ refused to invent a story to fill the silence.
I have reviewed a lot of garbage output in eight years of quantitative work. I have never been more relieved to see it.
The failure mode that should worry you in 2025 is not a model that breaks. It is a model that performs. Hand an LLM two empty fields and a mandate to analyze, and it will return something that reads like conviction โ clean headers, confident adjectives, a fabricated token supply table carried out to four decimal places. That is the difference between a tool and a liability. Yield without protocol is just delayed loss; the same is true of analysis without input.
Context
Here is the operational reality. Over the past eighteen months, a meaningful share of crypto funds โ my own team included โ began routing at least one layer of on-chain interpretation through language models. The pitch is seductive and the math is real. A junior analyst costs about $140,000 a year and reads forty projects a month. A well-prompted model reads four thousand and does not sleep. Institutional bridging between traditional finance and crypto now runs through these systems, and the ETF era only accelerated the dependency: when you are reconciling spot inflows against whale wallets at 09:30 New York time, you want automation, not a human squinting at a block explorer.
But automation has a tail. The report on my desk was the tail.
The structure of the failure is boring, which is exactly why it matters. Stage one was supposed to ingest a source document, parse it into discrete information points, and hand stage two a structured payload: title, source, thesis, named projects, timestamps, source quality. Stage one returned nulls across every field. Not an error. Nulls. The parser had executed, connected, and returned a well-formed document containing nothing. By the time stage two received it, the schema looked healthy and the content was a vacuum.
Any pipeline that checks only whether JSON parses will pass that payload straight into the reasoning layer. And a reasoning layer with no facts will still reason. That is the trap. The dangerous input is not the malformed one that throws an exception. It is the empty one that passes validation.
Core
Let me walk through what a correct system does, because the null report is not a bug to hide. It is a signal to read.
First, the taxonomy of inputs. There are three states a data payload can occupy, and conflating them is how desks lose money. State one: populated and verified. State two: populated and unverified. State three: empty. State three is not a lower-confidence version of state two. It is categorically different. An unverified claim can be graded โ one source, low quality, treat the thesis as speculation and size accordingly. An empty claim cannot be graded at all, because there is nothing in it to grade. The moment you let a model treat silence as a soft no-data flag and still produce a thesis, you have manufactured signal out of a hole. Speculation is noise; fundamentals are signal. Empty input is neither. Holes do not get filled by confidence, only by fabrication.
Second, the cost asymmetry. I wrote about this after the 2017 ICO cycle, and the arithmetic has not changed. A correct null costs you one missed trade โ maybe 2% of a quarter's alpha. An incorrect fill costs you the position you build on top of it. In 2020, when I ran a three-developer arbitrage desk between Uniswap V2 and SushiSwap, our average execution latency was 400 milliseconds and our edge decayed within eight weeks as MEV bots saturated the spread. We killed the strategy the day the numbers stopped working, not the day the losses started. That discipline came from one rule: never trade on data you cannot reconcile. If the ledger does not balance, you do not adjust the ledger. You stop.
Third, the recurrence problem. A model that hallucinates once will hallucinate again, because hallucination is not an accident. It is the path of least resistance โ the statistically cheapest way to satisfy a prompt asking for an answer. The probability that an LLM invents a token supply table when asked about an empty field is not small. It is the default behavior. In an internal test my team ran last quarter, we fed three models an empty payload with a standard nine-dimension analysis prompt. Two produced complete, internally consistent reports with plausible numbers. One fabricated a supply of 1,000,000,000 tokens with a four-year vesting schedule. The other invented a DeFi category label and then analyzed the yield it had just made up. Only the system with a hard null-gate returned N/A across the board. The lesson is not that models are dishonest. It is that they are eager to be useful, and usefulness without verification is the most expensive quality a tool can have.
Fourth โ and this is the part most desks miss โ the null report is itself data. The fact that stage one returned empty tells me something concrete: either source ingestion failed, or the source did not exist, or the schema mapping broke between stages. Those are three different problems with three different fixes, and none can be solved downstream. A desk that responds to a null by adding more analysis is doing what the 2022 algorithmic-stablecoin crowd did: doubling into a broken position and calling it conviction. Volatility is the tax on undiscerned capital. Empty data is the tax on undiscerned process, and it compounds faster.
Contrarian
The counter-intuitive part, and the reason I am publishing this at all.
The industry narrative right now is that AI will democratize on-chain analysis โ that any retail trader with a twenty-dollar subscription will soon read order flow like a fund. That narrative is half right and dangerously incomplete. What AI actually does, at scale, is accelerate whatever the operator already is. Give it to a disciplined desk and it multiplies discipline. Give it to a hype-driven one and it manufactures a thousand coherent-sounding theses an hour, each one a small coffin with a nice font. The market pays for clarity, not complexity. A model that generates complexity is a liability; one that enforces clarity is an edge.
This is why I trust the null report more than any bullish output I received last month. A system willing to say I have nothing has, by definition, not been trained to please me. A system that always answers is a system that cannot be trusted to flag the moment it stops knowing.
Notice where the real risk has migrated. We spend enormous energy auditing smart contracts for admin-key exposure and centralized sequencers. Correct. Layer2 sequencers remain effectively single nodes, and decentralized sequencing has been a slideware promise for two years. But the newest single point of failure on a modern desk is not the contract. It is the data pipeline feeding the model that reads the contract. An oracle that lies once poisons every position downstream. A parser that returns null and gets ignored does exactly the same thing, just more quietly, and under a compliance report that looks clean.
Takeaway
The report on my desk will not be published as analysis, because it is not analysis. It is a tripwire โ a machine confirming that when the input is empty, the honest output is empty too. The desk that treats that as a failure will keep hallucinating its way toward a margin call. The desk that treats it as a gate will keep its capital.
When your models next hand you conviction, ask one question before you size the position: what would this system return if I gave it nothing? If you do not know the answer, you do not know your edge. I trade the ledger, not the hype cycle โ and right now the ledger is telling me the most valuable output in crypto is a properly set gate.
